# Upgrade verification: 2026-09-08

The upgraded context plugin was checked against both the published Harness
release and the newer upstream source checkout. The plugin delegates execution
to the stock loop and uses durable surface replacements for conversational
context. Both stock session and agent-loop invariants remain enabled.

| Target | Executed evidence | Result |
| --- | --- | --- |
| Published `@deepseek-ai/dsh@0.1.3-alpha.2` | Packed plugin installed through `dsh plugin`; supported profile boot through Loader; four turns, five requests, real `recall_turn`, and JSONL close/resume | Passed, four surface replacements, zero agent errors |
| Upstream master `c389f96bf3a9b6807cb71ed6bdad5849be0df6d8` | Entire plugin suite with every Harness/vendor workspace import resolved to that checkout's source | 153 passed, zero skipped |

The verified plugin tarball was version `0.1.0`, with SHA-256
`251162b4130d211a303873fdc28107854df93632105daeb32ab780d04d04f74a`.
The verification host used Node `v22.22.3`, pnpm `11.7.0`, and macOS arm64.

The release source is commit `82a5fd61a7cf5c293cec4bdff68f455398d685e9`.
Between that commit and the verified master, upstream moved concrete inbox
storage from `dsh-agent` to the stock agent loop, changed inbox projection and
scope ownership, and added `FileSystem.readByteRange`. The upgraded plugin does
not own those changing runtime implementations.

## Repeat the packed check

Build the plugin, then pack and verify the current checkout:

```sh
pnpm run build
node scripts/validation/run-packed-smoke.mjs
```

To verify an existing tarball instead:

```sh
node scripts/validation/run-packed-smoke.mjs "$SLICE_PLUGIN_TARBALL"
```

The script creates a fresh temporary directory, installs the exact published CLI,
creates an isolated `DSH_HOME`, and uses the launcher's standard profile settings:
`nodeLinker: hoisted` and `autoInstallPeers: false`. It invokes the supported
`dsh plugin --profile slice-packed add` command, checks the composed Loader tree,
and boots that profile. It leaves the user's installed profiles unchanged.
Only persistence's declared `fs-ext` native build script is run; a native C++
toolchain is required. Package installation requires network access.

The fixture uses a deterministic LLM adapter. Settings, tool dispatch, recall,
the stock loop, surface derivation, invariants, and JSONL persistence use published
implementations. Every captured request is independently reconstructed from its
durable surface and header. The fixture checks unchanged runtime retention,
successful recall, history replacement, known event vocabulary, and an exact
persisted-prefix reload before continuing the session.

The printed evidence directory contains the verified artifact digest, summary,
captured requests, final events, persistence files, and each subprocess's output.
No API credentials are needed.

## Repeat the source check

Use a separate Harness checkout at the desired upstream revision. In that
checkout, install the required source dependency closure and build the one native
persistence dependency used by these tests:

```sh
pnpm install --filter @deepseek-ai/dsh-agent-loop... \
  --filter @deepseek-ai/dsh-session-persistence-jsonl... \
  --filter @deepseek-ai/dsh-tool-fs... --ignore-scripts --frozen-lockfile
npm --prefix node_modules/.pnpm/fs-ext@2.1.1/node_modules/fs-ext run install
```

From the plugin checkout, run:

```sh
node scripts/validation/run-master-tests.mjs "$HARNESS_SOURCE_CHECKOUT"
```

The script uses the checkout's complete `tsconfig.base.json` path map and its
standard-decorator transform. Harness imports all resolve to one source graph;
it does not mix selected master packages with published Harness artifacts. The
printed evidence directory records the exact upstream commit, test report, and
command output. The verified source worktree remained clean.

These checks establish execution, context, packaging, and persistence
compatibility on the tested host. They do not establish real-model completion
quality, total-token cost, or a cross-platform compatibility matrix. No real-model
benchmark or complete upstream application build was performed.